|
| 1 | +import sys |
| 2 | + |
| 3 | +import numpy as np |
| 4 | +import pytest |
| 5 | + |
| 6 | +import pyopencl as cl |
| 7 | +from pyopencl.tools import ( |
| 8 | + pytest_generate_tests_for_pyopencl as pytest_generate_tests, # noqa |
| 9 | +) |
| 10 | + |
| 11 | +import loopy as lp |
| 12 | +from loopy.kernel.dependency import ( |
| 13 | + add_lexicographic_happens_after, |
| 14 | + reduce_strict_ordering, |
| 15 | +) |
| 16 | + |
| 17 | + |
| 18 | +def test_no_dependency(): |
| 19 | + t_unit = lp.make_kernel( |
| 20 | + "{ [i,j] : 0 <= i, j < n}", |
| 21 | + """ |
| 22 | + a[i,j] = 2*i {id=source} |
| 23 | + b[i,j] = a[i,j] {id=sink} |
| 24 | + """, |
| 25 | + ) |
| 26 | + |
| 27 | + t_unit = add_lexicographic_happens_after(t_unit) |
| 28 | + t_unit = reduce_strict_ordering(t_unit) |
| 29 | + knl = t_unit.default_entrypoint |
| 30 | + |
| 31 | + assert len(knl.id_to_insn["sink"].happens_after) == 0 |
| 32 | + |
| 33 | + |
| 34 | +@pytest.mark.parametrize("img_size", [(512, 512), (1920, 1080)]) |
| 35 | +def test_3x3_blur(ctx_factory, img_size): |
| 36 | + ctx = ctx_factory() |
| 37 | + queue = cl.CommandQueue(ctx) |
| 38 | + |
| 39 | + hx, hy = img_size |
| 40 | + img = np.random.default_rng(seed=42).random(size=(hx, hy)) |
| 41 | + |
| 42 | + knl = lp.make_kernel( |
| 43 | + "{ [x, y]: 0 <= x < hx and 0 <= y < hy }", |
| 44 | + """ |
| 45 | + img_(i, j) := img[i+1, j+1] |
| 46 | + blurx(i, j) := img_(i-1, j) + img_(i, j) + img_(i+1, j) |
| 47 | +
|
| 48 | + out[x, y] = blurx(x, y-1) + blurx(x, y) + blurx(x, y+1) |
| 49 | + """, |
| 50 | + [ |
| 51 | + lp.GlobalArg("out", |
| 52 | + dtype=np.float64, |
| 53 | + shape=(hx, hy), |
| 54 | + is_output=True), |
| 55 | + lp.GlobalArg("img", |
| 56 | + dtype=np.float64, |
| 57 | + shape=(hx, hy)) |
| 58 | + ] |
| 59 | + ) |
| 60 | + |
| 61 | + knl = lp.fix_parameters(knl, hx=hx-2, hy=hy-2) |
| 62 | + |
| 63 | + bsize = 4 |
| 64 | + knl = lp.split_iname(knl, "x", bsize, inner_tag="vec", outer_tag="for") |
| 65 | + knl = lp.split_iname(knl, "y", bsize, inner_tag="for", outer_tag="g.0") |
| 66 | + knl = lp.precompute( |
| 67 | + knl, |
| 68 | + "blurx", |
| 69 | + sweep_inames="x_inner, y_inner", |
| 70 | + precompute_outer_inames="x_outer, y_outer", |
| 71 | + precompute_inames="bx, by" |
| 72 | + ) |
| 73 | + |
| 74 | + knl = lp.prioritize_loops(knl, "y_outer, x_outer, y_inner, x_inner") |
| 75 | + knl = lp.expand_subst(knl) |
| 76 | + |
| 77 | + knl = add_lexicographic_happens_after(knl) |
| 78 | + knl = reduce_strict_ordering(knl) |
| 79 | + |
| 80 | + _, out = knl(queue, img=img) |
| 81 | + blurx = np.zeros_like(img) |
| 82 | + out_np = np.zeros_like(img) |
| 83 | + for x in range(hx-2): |
| 84 | + blurx[x, :] = img[x, :] + img[x+1, :] + img[x+2, :] |
| 85 | + for y in range(hy-2): |
| 86 | + out_np[:, y] = blurx[:, y] + blurx[:, y+1] + blurx[:, y+2] |
| 87 | + |
| 88 | + import numpy.linalg as la |
| 89 | + assert (la.norm(out[0] - out_np) / la.norm(out_np)) <= 1e-14 |
| 90 | + |
| 91 | + |
| 92 | +if __name__ == "__main__": |
| 93 | + if len(sys.argv) > 1: |
| 94 | + exec(sys.argv[1]) |
| 95 | + else: |
| 96 | + from pytest import main |
| 97 | + |
| 98 | + main([__file__]) |
0 commit comments